Monday, July 23, 2007

Using graphics cards to do science

I'm stuck working from home for a second day thanks to a sudden-onset summer cold, but while I back-up all my data here is an interesting article from today's astro-ph: astro-ph/0707.2991.

Schive et al, at the National Taiwan University in Taipei, describe using the highly-parallel processors on modern GPUs (specifically a cluster of machines with dual nVidia GeForce 8800 GTX graphics cards) to perform astrophysical N-body simulations. A special library and software development kit called CUDA, developed by nVidia, allows the programmers to perform computation on the GPU and access the results instead of creating graphical output to a monitor.

Compared to the dedicated GRAPE processors that are designed to specifically address this kind of problem in hardware their PC+GPU cluster performs extremely well (both in performance per unit price, and in net teraflops).

This is very exciting, as the increasing capabilities of commodity PC hardware bode well for the future of scientific computing. Back in the mid-1990's when I started my PhD the department hard a few, very expensive, Sun Sparc and DEC Alpha workstation that were used by everyone through thin client terminals. By the time I finished my PhD commodity Pentium and Athlon PCs running linux offered equivalent or better performance per unit price, and people started having individual machines to themselves* (although the fastest CPUs in absolute terms were still the ultra-expensive Compaq [formerly DEC] Alpha's). Beowulf clusters appeared at the same time, and now clusters based on essentially PC hardware dominated the chart of the top 100 supercomputers.

Much of the growth in PC CPU power over the last decade has been driven by multiple factors: PC gaming, the growth of the Internet, the AMD vs. Intel competition, and also ironically the need for fast CPUs to counteract the growing bloat and slothfulness of Microsoft's Windows 98/XP/Vista series.

Now the increasing graphical demands of PC gaming, and the nVidia vs ATI (now part of AMD) market share competition may drive the rate of improvement in scientific computing for the next decade.

* Although the move to semi-individual machines rather than multi-user use was also driven at the by the poor network performance of NFS cross-mounted IDE disks compared to NFS mounted SCSI drives on the traditional workstations.

Post a Comment